BRENNEN

Industries / Media, arts and recreation / Media and entertainment / Video games

Art and animation

Artists and animators make what the player sees and hears: characters, places, movement, faces and voices, all built to be drawn many times a second while the game runs.

6

Milestones achieved, from 1989 to 2025.

1

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1989

    Motion traced from video

    Jordan Mechner made the runs, jumps and climbs in Prince of Persia move like a person's by rotoscoping, drawing animation frames over filmed motion. In 1985 he videotaped his 15-year-old brother running and jumping in a high school parking lot, then worked out how to fit that animation into an Apple II computer. Brøderbund published the game in 1989, and by Mechner's account it was converted to nearly every console and computer and sold 2 million copies. His 1984 game Karateka had already used rotoscoped animation.

    jordanmechner.comjordanmechner.com

    Credit

    Jordan Mechner; Brøderbund

  2. 2013

    Materials described in plain values

    At SIGGRAPH 2013, Brian Karis of Epic Games described the "physically based" material model of Unreal Engine 4, adapted from one Disney had used for the film Wreck-It Ralph. Artists describe each surface with four values (base color, metallic, roughness and cavity, the last for tiny cracks and seams) and the engine works out how light reflects from it. Karis wrote that Epic's artists expected "an enormous improvement to workflow and quality." Fortnite, then in production, had a deliberately unrealistic look that the new model could not represent, so Epic kept the old model as a switch for that game.

    LessonA new standard way of describing surfaces still needed a way out for a game whose art direction broke its rules.

    blog.selfshadow.com

    Credit

    Brian Karis and the Epic Games rendering team; Brent Burley, Disney

  3. 2014

    Scenery scanned from photographs

    In March 2014 The Astronauts described how they built the scenery of The Vanishing of Ethan Carter with photogrammetry, making 3D models from many overlapping photographs in Agisoft's PhotoScan software. A boulder took 26 photos and a quarry wall about 30 m (100 feet) high took 242. Raw scans held 2 to 20 million triangles each, and Andrzej Poznanski of the studio wrote that "you need to be really skilled at geometry optimization and retopology," rebuilding a model with far fewer triangles, before a scan can run in a game.

    LessonScanning moved the artist's work from modeling to cleanup rather than removing it.

    theastronauts.com

    Credit

    Andrzej Poznanski and The Astronauts

  4. 2020

    Lip sync in ten languages

    For Cyberpunk 2077, speech animation was generated for ten languages, some scenes mixing languages within one sentence, with tools from Jali that move a character's mouth, neck, brows and eyes from recorded audio and a tagged transcript. Animators could still adjust each result by hand on a standard face rig. The team described the work at SIGGRAPH 2020.

    history.siggraph.orgdoi.org

    Credit

    Pif Edwards, Chris Landreth, Eugene Fiume and Karan Singh of Jali, with Mateusz Popławski, Robert Malinowski and Sarah Watling

  5. 2025

    Actors win AI consent rules

    Video game performers in SAG-AFTRA, the US actors' union, were on strike against game companies from July 2024. In May 2025 the union filed a charge against Llama Productions, an Epic Games company, over an AI-generated Darth Vader voice in Fortnite. On 9 July 2025 members approved a new contract by 95.04% to 4.96%: companies must give notice and get consent before using AI to copy a performer's voice, face or movements, performers can suspend consent for new AI material during a strike, and pay rose 15.17% at once.

    LessonRules for AI copies of game performers came from a contract won after an 11-month strike, not from a law.

    graphicpolicy.comdecrypt.co

    Credit

    SAG-AFTRA's video game performers and negotiators

  6. 2025

    AI textures cost an award

    On 18 December 2025 the Indie Game Awards named Sandfall Interactive's Clair Obscur: Expedition 33 its Game of the Year and Debut Game, and two days later withdrew both. Sandfall's representatives had agreed at submission that no generative AI was used, which the awards' rules forbid; on the day of the ceremony Sandfall confirmed that AI-generated images had been used as placeholder textures, some of which shipped in the released game before a patch replaced them. The awards went to Blue Prince and Sorry We're Closed.

    LessonStand-in art made with AI can get past testing into a shipped game, and some juries count any use at all.

    engadget.comtribune.com.pk

    Credit

    Indie Game Awards; Sandfall Interactive

Watch for

  1. Faces animated from speech

    NVIDIA open-sourced Audio2Face in September 2025: models that turn recorded speech into lip sync and facial expression, a model that reads emotion from the voice, plugins for Autodesk Maya and Unreal Engine 5, and a framework for training new models. NVIDIA lists Codemasters, GSC Game World, NetEase and Perfect World Games among its users and quotes Survios and The Farm 51 on using it in Alien: Rogue Incursion Evolved Edition and Chernobylite 2: Exclusion Zone. It calls the model "industry-leading" but publishes no measure of its quality against hand animation or facial capture.

    Achieved whenAn outside study rates its output against hand-keyed or captured animation in blind tests, in more than one language.

    developer.nvidia.com

    Claimed by

    NVIDIA, September 2025

Open

  1. 01

    Rate AI lip sync blind

    A blind viewing test of speech animation made from audio alone. Record the same lines in five languages, animate one character's face three ways (with an audio-driven tool such as Audio2Face, with facial capture of the actor, and by hand), and have native speakers rate how well each clip's mouth matches the words without knowing how it was made.

    The brief
    Credit · unclaimed
  2. 02

    Time cleanup of AI props

    A timed test of how long game artists take to turn AI-generated 3D models into props a game can use, against modeling the same props from scratch. Generate 30 props, such as a crate, a lantern and a chair, with generators such as TRELLIS.2, Hunyuan3D and Meshy; hand them to artists with a fixed target (for example under 5,000 triangles, clean UV maps, which tell the engine how to wrap a texture, and textures that import into Unreal Engine or Godot); record their minutes and every fix.

    The brief
    Credit · unclaimed

Open 01 · the brief

Rate AI lip sync blind

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A blind viewing test of speech animation made from audio alone. Record the same lines in five languages, animate one character's face three ways (with an audio-driven tool such as Audio2Face, with facial capture of the actor, and by hand), and have native speakers rate how well each clip's mouth matches the words without knowing how it was made.

Why it is still open

NVIDIA's announcement of Audio2Face gives no quality measures, and the Cyberpunk 2077 talk describes a workflow, not a viewer study. A 2023 ArtsIT conference paper by Mingzhu Wei, Nicoletta Adamo, Nandhini Giri and Yingjie Chen compared four ways of animating faces, audio-driven among them, but not across languages. This research found no published viewer study that compares audio-driven game lip sync with hand animation or capture across languages.

developer.nvidia.comhistory.siggraph.orgapi.crossref.org

Built with what exists

NVIDIA's open Audio2Face models and their Maya or Unreal Engine 5 plugin; a free rigged character head; a phone or webcam facial-capture app; consenting voice actors in five languages; a speech recognizer such as ElevenLabs Scribe v3 to time each word (see the Note "Fewer word errors"); an online survey form for raters.

  1. First step

    Record 10 English and 10 Spanish lines, run them through Audio2Face, and try the rating form on 10 viewers.

  2. Done looks like

    50 lines in each of 5 languages, 3 versions of each, rated by at least 20 native speakers per language on a 5-point scale; clips, ratings and scripts published; the gap between audio-driven and hand-made versions reported for each language.

Open 02 · the brief

Time cleanup of AI props

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A timed test of how long game artists take to turn AI-generated 3D models into props a game can use, against modeling the same props from scratch. Generate 30 props, such as a crate, a lantern and a chair, with generators such as TRELLIS.2, Hunyuan3D and Meshy; hand them to artists with a fixed target (for example under 5,000 triangles, clean UV maps, which tell the engine how to wrap a texture, and textures that import into Unreal Engine or Godot); record their minutes and every fix.

Why it is still open

A 2026 survey by Jiafeng Wu and colleagues compared eight generators, including TRELLIS.2, Hunyuan3D 3.1 and Meshy AI 5, under the same rendering, and describes "a persistent gap" between their output and the production-ready standard, but reports no cleanup times. Other published comparisons found are vendor blogs and preference votes. This research found no timed study with working game artists.

christinebobby.github.ioarxiv.org

Built with what exists

TRELLIS.2 and Hunyuan3D (models free to download) and Meshy (a paid service); Blender, with a script that counts triangles, overlapping UVs and broken geometry; Unreal Engine or Godot to check each import; screen recording to log time.

  1. First step

    Generate 5 props with two generators, have one artist clean them up and model one of them from scratch, and log the times.

  2. Done looks like

    30 props, each made both ways by at least 2 artists; median minutes per prop for each route; the share of AI models that had to be remodeled from scratch; files, scripts and time logs published.